ICU EHR Data Gaps: Reconcile Pumps, Rethink Alerts

Predictive Reviews

Authors:João Brainer Clares de Andrade; Marconny Alexandre Oliveira de Medeiros Cavalcante; Thiago Luís Marques Lopes, et al.

Journal:Critical Care

DOI:10.1186/s13054-025-05677-0

ICU AI is only as safe as its data: a systematized review links EHR issues to 34% of ICU medication errors and finds pumps logging more dosing discrepancies than MARs (682 vs 321). Pharmacy leaders should prioritize pump-to-EHR reconciliation, missingness-aware models, and pharmacist-led real-time monitoring to safeguard high-risk infusions and CDS.

January 8, 2026

Quick Take

ICU electronic health record (EHR) data‑quality problems can materially affect medication safety: across included studies, 34% of reported ICU medication errors were attributed to EHR‑related issues (roughly one‑third assessed as potentially life‑threatening). Comparisons between smart‑pump logs and medication administration records (MARs) in available studies reported 682 device‑logged dosing discrepancies versus 321 documented in MARs, suggesting EHR documentation may under‑detect some device‑level dosing errors; these comparisons often derive from single‑center or non‑adult cohorts and require local validation.

For pharmacy leaders: prioritize device‑to‑EHR reconciliation and near‑real‑time pump→EHR comparison; triage pharmacist review of mismatches (vasopressors, narcotics, total parenteral nutrition); and require external and subgroup validation plus ongoing pharmacist oversight before integrating machine‑learning predictions into ICU care pathways.

Why it Matters

ICU pharmacy depends on EHRs, MARs, and device data to verify dosing, titrate high‑risk therapies, and support surveillance. When data are incomplete, delayed, or inconsistent across pumps, vitals, labs, and notes, true dose and timing changes can be obscured, leaving hidden discrepancies and patient safety risks.

Health systems increasingly use ML‑driven clinical decision support (CDS) for sepsis, acute kidney injury, and adverse event surveillance; variable ICU data quality over time and across sites can degrade predictions, increase alert burden, and misdirect scarce pharmacist attention. Under staffing constraints, weak data governance and lack of real‑time monitoring may compromise stewardship, medication safety, and CDS reliability.

What They Did

Conducted a systematized review of adult ICU and emergency department (ED) EHR data‑quality literature (2010–2024) across PubMed/MEDLINE, Embase, IEEE Xplore, ACM Digital Library, CINAHL, Web of Science, Google Scholar, DBLP, and the Cochrane Library using explicit inclusion/exclusion criteria.

Employed single‑reviewer screening with 10% double‑screening; 29 studies met eligibility and underwent full‑text extraction.

Extracted study setting, sample size, EHR system and data types, data‑quality metrics, machine‑learning approaches, and implementation details; used a pragmatic quality appraisal rather than formal risk‑of‑bias tools.

Performed a narrative synthesis and developed a taxonomy of ICU EHR data‑quality issues to derive implementation‑focused recommendations for monitoring and preparing ML‑ready datasets.

What They Found

Missingness: analyses of some large ICU datasets (e.g., MIMIC‑III) showed that certain variables had >80% missing; across several studies, >40% of predictors selected by feature‑selection methods were missingness indicators rather than measured values, with missingness patterns becoming more informative around ICU days 2–3.

Medication/documentation: multiple studies reported EHR‑related medication errors (cited pooled finding ~34% of events, with about one‑third assessed as potentially life‑threatening). Comparisons of smart‑pump logs vs. MARs in available studies found more discrepancies logged by pumps (682 vs. 321), concentrated in vasopressors, narcotics, and TPN; these comparisons were often single‑center and, in some cases, from neonatal settings, so direct generalization to all adult ICUs is limited.

Temporal degradation and generalizability: high‑frequency vital‑sign data quality deteriorated during stays (quality issues tended to emerge by ~60–75% into average length of stay depending on the signal); only ~53–67% of stays maintained acceptable vital‑sign coverage throughout admission. Sepsis prediction models reported internal AUCs in the mid‑0.7s but dropped to ≈0.63 on external validation in several studies.

Mitigation signals: some automated detection and imputation approaches reported high, task‑specific performance (examples include reported anomaly‑detection metrics ≈99.2% accuracy and imputation evaluations reporting ~84.6% 'success' with ~35.6% reduction in missingness). These results are study‑ and task‑specific and should be interpreted cautiously pending broader, local validation.

Takeaways

Treat the EHR as an imperfect source, not ground truth. Pharmacists should adopt a 'trust but verify' stance: expect documentation to lag device logs, particularly for vasopressors, narcotics, and TPN, and validate critical dosing from device records where possible.

Use ML outputs as triage signals rather than definitive actions. Require transparent handling of missingness (especially around days 2–3), external and subgroup validation, and monitoring for performance drift with longer stays or new sites.

Implement quality gates rather than piling on unfiltered alerts. Where feasible, deploy real‑time anomaly detection and device integration, define prioritization and escalation rules, and specify who (pharmacist or clinical team) reviews and adjudicates outliers.

Monitor equity implications. Measurement frequency and data availability can vary by demographic group and care setting; insist on local fairness evaluation, continuous monitoring, and retention of pharmacist clinical judgment in decision loops.

Strengths and Limitations

Strengths:

Comprehensive, multidisciplinary search across major databases guided by expert‑derived questions, enabling a focused synthesis of ICU EHR data‑quality literature.; Narrative synthesis produced an implementation‑oriented taxonomy and practical recommendations with transparent methods (extraction fields, search strings, supplementary material) to aid reproducibility.

Limitations:

Single‑reviewer screening (10% double‑screened) and exclusion of grey literature and foundational dataset creation papers may introduce selection bias and limit completeness.; Pragmatic quality appraisal and heterogeneity of included studies precluded formal risk‑of‑bias assessment or meta‑analysis, limiting quantitative generalizability of pooled estimates.; Some headline quantitative signals (e.g., smart‑pump vs. MAR discrepancy counts, very high anomaly‑detection accuracy, imputation 'success' rates) derive from single‑center, neonatal, or task‑specific studies with variable sample sizes and may not generalize to adult, multicenter ICU settings without local validation.

Bottom Line

Safe ICU AI depends on reliable operational 'plumbing': invest in device‑to‑EHR reconciliation, missingness‑aware modeling, rigorous external and subgroup validation, and pharmacist‑led real‑time monitoring and adjudication prior to large‑scale clinical adoption.